audio-transcribe

Pass the YouTube video ID or URL to the transcribe function, which uses yt-dlp to extract the audio as mp3 and sends it to OpenAI's transcription API. The result is returned as plain text along with chunk count and audio duration.

First transcode the audio to mono 16kHz mp3 at 32kbps with ffmpeg, which compresses it roughly 10x. If it still exceeds the limit, split it into time-based segments, transcribe each separately, and concatenate the text results.

gpt-4o-transcribe is OpenAI's more accurate transcription model with lower word error rate, but it outputs plain text only. Use whisper-1 when you need SRT or VTT output with segment-level timestamps.

No, OpenAI's transcription API does not distinguish between speakers. The output is a single continuous text stream without speaker labels, so multi-speaker attribution requires a different tool.

The transcription pipeline requires ffmpeg for transcoding and segmentation and ffprobe for duration measurement, both on the system PATH. Install them with a package manager such as brew install ffmpeg on macOS.

Pass a prompt parameter containing the correct spellings of names and terms as context to bias the transcription model. This mitigates common ASR errors like mishearing brand or person names.